CE-RS-SBCIT: A novel channel-enhanced hybrid CNN–transformer with residual, spatial, and boundary-aware learning for brain tumor MRI analysis

Abstract

Brain tumors are among the most life-threatening diseases, where early detection and accurate classification are critical for effective diagnosis and treatment planning. Although deep learning-based computer-aided diagnosis systems have shown remarkable progress, conventional convolutional neural networks (CNNs) and Vision Transformers (ViTs) still face challenges such as limited global-local feature integration, sensitivity to structural heterogeneity, and high computational cost. To address these limitations, this study proposes a novel hybrid framework, termed CE-RS-SBCIT, which integrates residual and spatial learning with transformer-driven contextual modeling. The proposed framework architecture combines five core components: (i) a smoothing and boundary-based CNN-integrated Transformer (SBCIT) for enhanced local-global representation learning, (ii) residual and spatial CNNs modules for capturing fine-grained textures and structural variations, (iii) a multi-scale channel attention and squeezing (MSCAS) mechanism for adaptive feature weighting, (iv) a channel enhancement (CE) strategy for enriched feature fusion, and (v) a pixel-level spatial attention module for improved inter-class discrimination. Extensive experiments on publicly available magnetic resonance imaging (MRI) datasets demonstrate that the proposed framework achieves superior performance, with 98.30% accuracy, 98.08% sensitivity, 98.43% precision, and an F1 Score of 98.25%. Comparative and ablation analyses further confirm the effectiveness of the proposed architecture in capturing both heterogeneous and homogeneous tumor characteristics while maintaining computational efficiency. These results indicate that CE-RS-SBCIT provides a robust and generalizable solution for brain tumor classification and has strong potential for deployment in clinical decision support systems.